Download Automatic Detection and Correction for Chinese Misspelled Words

Automatic Detection and Correction for Chinese Misspelled Words
Using Phonological and Orthographic Similarities
Tao-Hsing Chang
Department of Computer Science
and Information Engineering, National
Kaohsiung University of Applied Sciences
[email protected]
Hsueh-Chih Chen
Department of Educational Psychology
and Counseling
National Taiwan Normal University
[email protected]
Yuen-Hsien Tseng
Information Technology Center
National Taiwan Normal University
[email protected]
Jian-Liang Zheng
Department of Computer Science
and Information Engineering, National
Kaohsiung University of Applied Sciences
[email protected]
‘代表’ is mistakenly written as ‘伐表’ because
the orthographic of characters ‘代’ and ‘伐’ are
quite confusing. As a result, it may work to identify the possible misspelled words within sentences by phonological similarity and orthographic similarity between two characters.
The purpose of the study is to propose a
method to detecting and correcting misspelled
words in sentences. The proposed method does
not rely on the collection of similar words, but
based on the following assumption. Supposing
there was no misspelled word in sentences, ideal
word segmentation method could divide sentence
into serial correct words. However, if there was a
misspelled word, the segmentation could separate words containing misspelled character by
serial characters. For instance, sentence ‘我們都
喜歡學佼’ will be segmented into ‘我們 都 喜歡
學 佼’ due to the fact that ‘學佼’ cannot be
found in the dictionary, thus segmenting ‘學’ and
‘佼’ respectively.
By the observation mention above, a sentence may include several character sequences
consisting of two or more than two characters,
denoted as sentential fragments. Each character
in fragments may be the wrong part of a misspelling word while other characters are the correct part of the word. Hence, for each character
treated as correct part of a misspelled word, the
proposed method picks up the words containing
the character. The words will be denoted as
“candidate words”. On the other hand, all characters in the fragment may be single-character
words. The sentential fragment referring to candidate words is called “original string” in this
Abstract
How to detect and correct misspelled words
in documents is a very important issue for
Mandarin and Japanese. This paper uses phonological similarity and orthographic similarity co-occurrence to train linear regression
model. Using ACL-SIGHAN 2013 Bake-off
Dataset, experimental results indicate that the
detection F-score, error location F-score of
our proposed method for Subtask 1 is 0.70
and 0.43 respectively, and the correction accuracy of the proposed method for Subtask 1
is 0.39.
1
Introduction
How to automatically detect and correct
misspelled words in documents is a very important issue. It is not an easy task for programs
to spot misspelled words automatically. In English sentences, words are separated by space,
thereby leading to the result that it is not difficult
to distinguish if there are characters with nonexisting orthography and unknown words. However, Chinese sentences are constructed by successive single-character, and a word could consist of one character or more. As a result, it is
difficult to identify whether a character is a part
of a misspelled word or not.
Based on our observation, misspelled words
mainly occur as the following cases: phonological similarity and orthographic similarity. For
example, word ‘已經’ is mistakenly written as
‘以經’ due to the fact that characters ‘已’ and
‘以’ are pronounced as ‘yi’. In addition, word
97
Proceedings of the Seventh SIGHAN Workshop on Chinese Language Processing (SIGHAN-7), pages 97–101,
Nagoya, Japan, 14 October 2013.
paper. By calculating the probability of candidate
words and original strings, the proposed method
can determine whether the original strings contain misspelled words or not and correct the
words. Next section will address the details.
2
the dictionary can be listed. In this study, these
words are called candidate word while the string
is called original string.
Linear regression prediction model was
used to determine whether an original string
should be replaced with a candidate word or not.
Three parameters between candidate word and
original string are used in linear regression model as an input. The values of three parameters are
respectively called similarity, the probability of
character co-occurrence, and the probability of
POS co-occurrence. The values are utilized as
the input in linear regression formula, and then
the probability of misspelled words in original
string can be obtained. If several candidate words
are predicted as the correct words of original
string by prediction model at the same time, the
word with highest score is treated as correct
word.
The similarity between candidate word and
original string is the average between phonological similarity and orthographic similarity. The
Mandarin phonetic code was employed to compute phonological similarity. High similarity
means that the two characters are easily represented as misspelled words for each other. On
the other hand, radical structures are utilized to
determine spatial structure of two fonts. Two
characters with higher graphic resemblance easily represents as misspelled words. In the following sections, the detail of similarity will be addressed.
Related Works
Chang (1995) proposed detecting technique
of Mandarin misspelled word. Although it was
able to find out misspelled words, there were
some defects needed to be improved. For example, too much False Alert, long detection time,
not able to refer to entire paragraph. Ren et al.
(2001) utilized rule-based with linguistic model
to detect mistakes. Although it was not very efficient, it was a new concept at the time. Lin et al.
(2002) focused on misspelled words occurred in
Cangjie input method and put forward a detecting system. Huang et al. (2008) designed a correcting system for wrong phonological words
which built up similar phonological word collection for every single word. The correcting system
also used bi-gram linguistic model to position the
misspelled word, and replaced it with the most
likely fit word.
Afterwards, there were many proposes under different circumstances. For example, Chen
(2010) following previous studies, he amended
detecting templates in order to automatically
generate positive and negative knowledge corpus
by Using Template and Translate modules to
correct sentences. And final correction was conducted by part-of-speech Language Model to
improve the accuracy probability of misspelled
word correction.
3
3.1
Candidate Words
For each sentence, it is segmented into
words and the part-of-speeches of words are
tagged by WeCAn system (Chang et al., 2012).
Based on the assumption mentioned earlier in
this paper, words which contain misspelled
words can result in consecutive single-character
string. Hence, the model will identify all words
contained in consecutive single-character string
from dictionary. As seen from Figure 1, a sentence ‘人生又何償不是如此’ is segmented into
‘人生_又_何_償_不是_如此’. ‘償’ is a misspelled words of ‘嘗’ in this sentence, ‘何’ and
‘償’ are thus segmented respectively. Followed
by, for the string with successive single-character
‘又何償’, the system will select the candidate
word from each character in the string. For the
‘何’ in string, the proposed method identifies ‘何’
as the second word and its length less than or
equal to 3 in the dictionary, such as ‘任何’, ‘如
何’ and so on. Additionally, the word ‘何’ identi-
Methods
Chang et al. (2012)’s approach is refined as
the algorithm for automatically correcting misspelled words in this paper. They observed that
there is a specific phenomenon when misspelled
words occur. They envisaged that there was no
misspelled words in sentences, ideal tokenization
system would divide sentence into correct vocabulary combinations. However, if there is a
misspelled word, the system would segment the
majority of vocabulary contained misspelled
words by means of single-character formation.
According to this property, existing misspelled words was assumed to appear in a string
formed by successive single-character words in
this paper. As a result, for a string including two
or more than two single-characters, the words
which contain some characters in a string from
98
and ‘ㄎ’, ‘ㄓ’ and ‘ㄗ’, ‘ㄔ’ and ‘ㄘ’, or ‘ㄕ’
and ‘ㄙ’.
2. If there are identical finals, the score will increase one point. The score will increase 0.5
point for finals of two characters which are ‘ㄣ’
and ‘ㄥ’.
3. If there are identical medial in two characters,
the score will increase one point.
4. If the tones are consistent in two characters,
the score will increase one point.
5. Phonetic similarity between two characters
can be obtained by dividing the similarity score
by 4.
Phonological similarity between candidate
word and original string is the average of phonological similarity of all characters in the candidate word. For instance, given candidate word
‘應該’ corresponding to original string ‘因該’,
phonetic symbol of characters ‘因’ and ‘應’ is
‘ㄧㄣ’ and ‘ㄧㄥ’ respectively. Hence, the phonological similarity is (1+1+0.5+1)/4 = 0.875.
The similarity between the same two characters
‘該’ is one. Therefore, phonological similarity
between candidate word ‘ 應 該 ’ and original
string ‘因該’is (1+0.875)/2=0.9375.
fied as the first word and its length less than or
equal to 2 is also the candidate word, such as ‘何
必’, ‘何嘗’and so on.
All candidate words will be compared to correspondent original string with their phonological similarity and orthographic similarity. In Figure 1, the phonological similarity and orthographic similarity between the character ‘任’ in
candidate word ‘任何’ and character ‘又’ in original string will be computed. The similarities
determine whether ‘任何’ is needed to be further
analyzed.
Figure 1. An Example of Candidate Words and
an Original String.
3.3
3.2
Orthographic Similarity
Measurement of orthographic similarity in
this paper is based on the method proposed by
Chang et al. (2012). The measurement first disassembles two characters into a set of basic components and compare the differences between the
two using Chinese Orthography Database proposed by Chen et al. (2011). There are 446 basic
constituents in the database, and each unit of a
character is linked by their spatial relations.
There are 11 types of spatial relations, such as
vertical combination and horizontal combination.
Through the database, a Chinese character
can be converted into a series of branch-like
structure consisting of parts and combination
relations. The structure is called the constituent
structure. Figure 2 shows the constituent structure of the Chinese character ‘查’ in which ‘-’
represents horizontal combination, and ‘木’, ‘曰’,
‘一’ are constituents.
In the constituent structure of a character,
nodes represent combination relations while
leaves represent constituents. Every relation and
constituent in the whole structure has a level representing the position as well as a weight representing the strokes for that constituent. The
weight for each relation denotes the total number
Phonological Similarity
Mandarin phonetic symbols are used to
evaluate phonological similarity. There are 37
symbols in Mandarin phonetic symbols dividing
into initial (ㄅ/b/,ㄆ/p/,ㄇ/m/), medial (ㄧ /yi/,ㄨ
/wu/,ㄩ/yu/), final (ㄚ/a/,ㄛ/o/,ㄜ/e/) and five
tones. Chang et al.(2010) mentioned that some
Chinese phonetic alphabets have identical articulation method and speech position, whereas articulation confusability is the causes of misspelled words. For example, symbols ‘ㄅ’ and ’
ㄆ’ have similar speech position ; symbols ‘ㄕ’
and ‘ㄙ’ have identical articulation; symbols ‘ㄣ’
and ’ㄥ’ in final category belongs to a confusable
articulation set.
This paper compares two characters with
their Mandarin phonetic symbols of its initial,
medial, final and tone respectively to measure
phonetic similarity. The rules for comparison are
as follows:
1. If there are identical initial, a similarity score
will achieve one point. The score will achieve
0.5 point for initials of two characters which are
‘ㄅ’ and ‘ㄆ’, ‘ㄉ’ and ‘ㄊ’, ‘ㄋ’ and ‘ㄌ’, ‘ㄍ’
99
of strokes for all relevant constituents. Chang et
al.(2012) used the level and weight of constituent
structure of two characters to calculate the degree of orthographic similarity.
obtained by dividing the PCC of characters in
candidate word by that in original string. The
ratio of the probability of part-of-speech cooccurrence (RPPC) can be obtained by the same
methods. The higher the two values, the possible
the misspelled words occur in original string. As
a result, both ratios will be two inputs for prediction model.
3.5
Figure. 2. Constituent Strucutre in the Chinese
Character ‘查’.
This paper modifies measuring formula for
similarity in previous study Chang et al. (2012).
In previous formula, only the similarity between
two identical constituents is scored. It is suggested that two similar constituents should increase
the score. Therefore, this paper uses the stroke
information in the database for constituents to
calculate the similarity between two constituents.
For example, stroke information of constituents
‘已’ and ‘己’ in the database is as follows:
‘已’ :[{口 2},{一}~(1:9@9),{乚}~(2:0@2)]
‘己’: [{口 2},{一}~(1:9@9),{乚}~(2:0@0)]
It is noted that most constituents for Chinese
characters ‘已’and ‘己’ are very similar which
would receive a score close to 1 in similarity
measure.
3.4
Probability of Co-occurrence
In large corpuses, some specific characters
may have high frequency to be adjacent to another character, and this is called co-occurrence.
The probability of co-occurrence between two
characters is called probability of character cooccurrence (PCC). If a sentence has misspelled
words, the PCC among characters in the sentence
should be lower than the sentence which has no
misspelled words. Hence, if the PCC of character
in the candidate word is great higher than that in
original string, misspelled words may occur in
the sentence. In addition, there exists cooccurrence in part-of-speeches. The probability
of co-occurrence between two characters is
called probability of character co-occurrence
(PPC).
Bi-gram method is utilized in both this paper and the previous study to calculate the PCC
of characters in candidate words as well as that
in original string. Ratio of PCC (RPCC) can be
Prediction Model
This paper adopts linear regression formula to
be the prediction model. For a candidate word
and correspondent original string, the values of
three inputs can be obtained by using approaches
in subsection 3.2 to 3.4. Candidate words and
correspondent original string in training data are
utilized in this paper to compute each regression
coefficient in formula 1.
(1)
For any set of candidate words and correspondent original string, three parameters are
substituted into formula 1 to obtain y. The y value represents the probability of the original string
within misspelled words. Based on the values of
y obtained from training data, a threshold can be
set. If y is higher above the threshold, there are
misspelled words in original string. If y is lower
than threshold value, there are no misspelled
words in original string.
4
Experiments
This paper use the data provided by ACLSIGHAN 2013 Bake-off to conduct performance
evaluation. The data is divided into two sets
called ‘dry run’ and ‘final test’ while the evaluation includes two tasks sub-task 1 and sub-task 2.
Each set consist of two subsets which is employed to evaluate the performance of methods
for two tasks respectively. In dry run, sub-task 1
and sub-task 2 each use 50 example sentences for
testing. In final test, sub-task 1 and sub-task 2
each use 1000 example sentences for testing. The
purposes of sub-task 1 and sub-task 2 are to respectively evaluate the performance of error detection and error correction of methods.
Table 1 presents the evaluation result of Subtask 1 in our proposed method, denoted as
KUAS-NTNU, and results of sub-task 2 are
shown in Table 2. Since SIGHAN has not reported the F-score for sub-task 1 in dry run, Table 1 does not show the detection F-score and
error location F-score of dry run.
100
False-alarm Rate
Detection Accuracy
Detection Precision
Detection Recall
Detection F-score
Error Location Accuracy
Error Location Precision
Error Location Recall
Error Location F-score
Dry Run
0.23
0.80
0.50
0.90
0.76
0.39
0.70
-
Final Test
0.23
0.79
0.61
0.82
0.70
0.69
0.38
0.51
0.43
Table 1. The Performance of KUAS-NTNU System for Subtask 1.
Location Accuracy
Correction Accuracy
Correction Precision
Dry Run
0.30
0.28
0.36
Final Test
0.44
0.39
0.51
Top University Project” of National Taiwan
Normal University (NTNU), sponsored by the
Ministry of Education, Taiwan. We are also
grateful to the support of International ResearchIntensive Center of Excellence Program of
NTNU and NSC, Taiwan, R.O.C., under the
grant 101WFA0300229.
References
Chang C.-H. 1995. A New Approach for Automatic
Chinese Spelling Correction. Proceedings of
Natural Language Processing Pacific Rim Symposium’95, 278-283.
Chang C.-H., Lin S.-Y., Li S.-Y., Tsai M.-F., Liao H.M., Sun C.-W., and Huang N.-E. 2010. Annotat-
ing Phonetic Component of Chinese Characters Using Constrained Optimization and Pronunciation Distribution. International Journal of
Table 2. The Performance of KUAS-NTNU System for Subtask 2.
5
Discussion
Methods suggested by previous studies often rely on data collected from confusable character sets. Although corresponding characters in
the set are high in similarity and can be easily
confused, they could not be assessed correctly if
they are not from the confusable sets. Our proposed methods calculate phonological similarity
and orthographical similarity between misspelled
words and original string, which are not restricted by confusable sets. The proposed method can
still obtain a reliable estimation by other parameters with characters in low similarity.
Some issues and works could be explored
and developed in the further. First, this study
only examines characters with misspelled words.
The detection and correction of single-character
misspelled words only rely on simple rule-based
approaches. It results in many single-character
misspelled words cannot be extracted. Second,
collections of unknown words in sentences are
often considered having misspelled words which
might cause a decrease in system accurate rate
but an increase in false-alarm rate. Differences
analysis for unknown words and misspelled
words are issues that must be dealt with in future
research.
Acknowledgments
This research was partially supported by
National Science Council (NSC), Taiwan, under
the grant NSC 100-2511-S-003-053-MY2 and
NSC 102-2511-S-151-002, and the “Aim for the
Computational Linguistics and Chinese Language
Processing, 15(2):145-160.
Chang T.-H., Su S.-Y., and Chen H.-C. 2012. Auto-
matic Correction for Graphemic Chinese Misspelled Words. Proceedings of ROCLING 2012,
125-139.
Chang T.-H., Sung Y.-T., & Lee Y.-T. 2012. A Chi-
nese word segmentation and POS tagging system for readability research. Proceedings of
42nd SCiP.
Chen H.-C., Chang L.-Y., Chiou Y.-S., Sung Y.-T.,
and Chang K.-E. 2011. Chinese Orthography
Database and Its Application in Teaching
Chinese Characters. Bulletin of Educational
Psychology, 43:66-86.
Chen, Y.-Z. 2010. Improve the Detection of Im-
properly Used Chinese Characters with Noisy
Channel Model and Detection Template.
Chaoyang University of Technology, Taiwan,
R.O.C.
Huang C.-M, Wu M.-C., and Chang C.-C. 2008. Er-
ror Detection and Correction Based on Chinese Phonemic Alphabet in Chinese Text.
World scientific publishing company, International
Journal of Uncertainty, Fuzziness and KnowledgeBased Systems, 16(1):89-105.
Lin Y.-J., Huang F.-L., and Yu M.-S. 2002. A Chinese Spelling Error Correction System. Proth
ceedings of the 7 Conference on Artificial Intelligence and Applications.
Ren F., Shi H., and Zhou Q. 2001. A hybrid ap-
proach to automatic Chinese text checking
and error correction. Proceedings of IEEE SMC,
1693-1698.
101